A compression method is proposed that takes advantage of a powerful property of acoustic unit inventories: In the appropriate acoustic space, units that share a (context-dependent or -independent) phoneme label must be close to a vector phoneme template associated with the phoneme. The method approximates units by interpolation between templates. The interpolation operation involves two asynchronous weight functions operating on the template. One is associated with spectral peak locations, the second with spectral balance. This enables approximating transitions such as [i:]→a[v], in which formant movement precedes frication onset. The algorithm guarantees smooth concatenation points.